3 ms·
> It implies to me that there's still a little bit of the characters in the dataset, but not very much. I suspect there are a lot of the characters in the trai
by sangnoir 4y ago
> It implies to me that there's still a little bit of the characters in the dataset, but not very much.
I suspect there are a lot of the characters in the training data, but (postulating) there might be a post-generation adversarial selection step where anything that is recognizable as a someone else's IP gets filtered out of the results, so anything sufficiently Pikachu-like won't be shown, leaving behind the vague yellow monsters.
- jamilton 4y agoI would think attempting to filter it out after generation would be harder and less effective.
- sangnoir 4y agoML image recognition predates diffusion models and is more mature - I'd be shocked if it's not employed in some form. Such a system would be more robust against incidental infringement of IP not mentioned in the prompt. e.g. A prompt for "Theme park" won't create Mickey Mouse in the background